| In the process of air-to-air combat confrontation,the attitudes and distances of the targets may change dramatically in a short time,and the accurate recognition of all the attitudes of the targets in the instantaneously large maneuver is the key to ensure real-time and accurate attack.Two-dimensional image information presented by three-dimensional information is continuously changing in airborne and missile-borne imaging equipment,therefore,grasping and utilizing effective information of the three-dimensional targets are important means for three-dimensional targets recognition.With the development of deep learning technology in various fields,especially the breakthrough of target recognition technology to extract high-level feature information,it provides an opportunity to furtherly improve the autonomous performance of combat weapons.Two-dimensional video image sequences presented from three-dimensional maneuvering target in the course of maneuvering change are adopted in this paper.As well as the spatiotemporal characteristics of the targets are used for reverse deduction and perception of the three-dimensional information.In order to study the three-dimensional target recognition technology based on deep learning method and spatiotemporal information perception,the specific work of this paper is summarized as follows:Firstly,considering the high redundancy of the original video image sequence data acquired by image acquisition equipment,which is not suitable to the acquisition of effective information of the targets,a key-frames extraction algorithm based on Capsules Network and self-attention mechanism is proposed.The key frames which can represent the vital information in the sequences are selected based on internal change mechanism of the sequence to form the key-frames sequence,and can also effectively reduce the redundancy of the sequence under the condition that the change characteristics of the original sequence of three-dimensional moving objects remain basically unchanged.Secondly,in order to recognize the three-dimensional target utilizing the spatiotemporal information in the process of maneuvering,in the meanwhile,considering the influence of different coupling degree of temporal information and spatial information,three-dimensional target recognition model structures based on spatiotemporally loose and tight coupling convolution neural network are proposed respectively.Spatiotemporally loose coupling model extracts the spatial and temporal features of the target separately,then fuses them to form spatiotemporal features to express the state of the three-dimensional moving targets;whereas the spatiotemporal tight coupling model takes use of sequential modeling to form spatiotemporal features to represent the targets.In order to furtherly increase the recognition of different attitude of three-dimensional targets,the distance measurement loss function is introduced after the spatiotemporal fusion process to effectively improve the aggregation of all attitudes of three-dimensional targets.Finally,eight different kinds of three-dimensional moving aircrafts are selected as experiment targets,and the detailed experimental analyses of spatiotemporally loose and tight coupling convolution neural network are carried out and the comparisons between them are also conducted.The experimental results show that the proposed spatiotemporally loose and tight coupling convolutional neural network models have the good performance for three-dimensional moving targets recognition.Similar attitudes of different targets as well as different attitudes of the same target can be accurately recognized.And at the same time,the reliability of key-frames extraction and the computational speed of the models are also verified. |
